1347
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
5. MODEL INPUT DATA
The Tb information collected from the aircraft, in its raw form,
includes latitude and longitude using Geocentric Datum of Aus
tralia 1994 (GDA 94) coordinates, brightness temperature value
(H-polarized, TbH and V-polarized, TbV), altitude and beam ID.
The altitude is used to determine the ground resolution required
and the beam ID is used for correcting to a common incidence
angle. For this study, the incidence angle is corrected to +/-
38.5° as this is a typical value for many satellite systems. The
medium resolution mapping with a flight altitude above sea
level (ASL) between 1050m to 1270m at Roscommon focus
farm was used. This results in a nominal ground resolution of
250m. The Tb data and the ground sampling soil moisture data
used are first georeferenced to the same coordinate system
(Universal Transverse Mercator, UTM). A regular grid was
created as the reference grid for the data. These grids divide the
area of interest into 250 x 250m square cells. Each cell of the
grid is next assigned a value of H-polarized brightness tempera
ture (TbH) and V-polarized brightness temperature (TbV) by
averaging all the points falling into each cell (Figures 2 and 3).
0 500 1,000 Meters
1 i i i i
Brightness Temperature(K)
m m
Figure 2. Aggregated H-Polarized brightness temperature at
Roscommon on 8 th November, 2005 using 250x250m grid cell.
Figure 3. Aggregated V-Polarized brightness temperature at
Roscommon on 8 th November, 2005 using 250x250m grid cell.
6. TESTING AND RESULTS
The Roscommon data are first divided into three different
groups or classes: low, medium and high, according to the
maximum and minimum value of the TbH data. This is to
ensure that the final data is distributed throughout the spatial
location and is not gathered only for a certain TbH range. For
each of the classes, the data are randomly divided into 60% for
training, 30% for validation and 10% for testing the trained
network. The validation set is used to stop the training of the
NN when the NN begins to overfit the data. The test dataset is
not used during the training and validation processes of NN
construction but is used subsequently to test the trained NN. A
general schematic of the division of the data is shown in Figure
4. The training, validation and testing set each contain values
from the low, medium and high classes. During the training and
validation processes, 10-fold cross validation is carried out
whereby the training set and validation set are combined.
During each run of the training process, a subset of this data
will be used for validation while the remaining data will be
used for training. The network is trained and validated using a
basic ANN that uses backpropagation with gradient descent.
The bias, layer weights and output weights of the network when
it produces the lowest Root Mean Square Error (RMSE) for
both the training and validation sessions are obtained. These
values are then used for the training, validation and testing of
the other backpropagation training algorithms using MATLAB.
This means that these ANN starts from a good configuration.
Table 2 shows the result of each of the backpropagation training
algorithm.